This file is automatically generated. It contains the descriptions of each uploaded file as provided by the user.

• 0.4_no_nlcc_DSH.zip: The semicore-based pseudopotentials have been purposely generated for this present study, and used for DSH band-gap calculations. The PBE pseudopotentials are generated using exactly the same parameters as for the PseudoDojo ONCVPSP-v0.4 except that nonlinear core correction is disabled. When multiple pseudopotentials are available for a single element, the most stringent one (i.e., with semicore states and a "high" suffix in the filename) are used in the present study.
• 0.4_fr_no_nlcc_DSH+SOC.zip: The semicore-based pseudopotentials have been purposely generated for this present study, and used for DSH+SOC band-gap calculations. The PBE pseudopotentials are generated using exactly the same parameters as for the PseudoDojo ONCVPSP-v0.4 except that nonlinear core correction is disabled. When multiple pseudopotentials are available for a single element, the most stringent one (i.e., with semicore states and a "high" suffix in the filename) are used in the present study.
• 246_structures.zip: The 246 structures used in this work to construct the DSH band-gap dataset, with all provided structures fully relaxed using the PBE functional.
• datasets_246_DSHgap+dielectric_constant.xlsx: The calculated band gaps with the functional of PBE and DSH including the spin-orbit coupling effects, as well as the corresponding dielectric infinity for the 246 materials in our dataset. The calculated band gaps determined using the PBE and DSH functionals, incorporating spin-orbit coupling effects, as well as the corresponding dielectric infinity values, for the 246 materials in our dataset
• ML_gap_figure3a&b.xlsx: The band gaps obtained using DSH+SOC vs. using the machine learning model [i.e. equation 1 (3D@1C) in the main text], as well as the adopted input features. These results are illustrated in figures 3a and 3b in the main text.
• ML_gap_figureS5a&b.xlsx: The band gaps obtained using DSH+SOC vs. using the machine learning model [equation S1 3D@4C(dielectric_infinity)] in the Supporting Information, as well as the adopted input features. These results are illustrated in figures S5a and S5b in the Supporting Information.
• Database_I_screening.xlsx: The 3D@1C machine learning model applied to Materials Project (Database I), including various information such as the screening criteria, predicted band gaps, DSH+SOC band gaps, and their respective input features.
• Database_II_screening.xlsx: The 3D@1C machine learning model applied to Database I, including various information such as the screening criteria, predicted band gaps, DSH+SOC band gaps, and their respective input features.